Обсуждение: Inaccurate statement about log shipping replication mode
The following documentation comment has been logged on the website: Page: https://www.postgresql.org/docs/17/warm-standby.html Description: Hello, The documentation page about Log-Shipping Standby Servers after describing that there are file-based log shipping and record-based log shipping (streaming replication) states: "It should be noted that log shipping is asynchronous, i.e., the WAL records are shipped after transaction commit.". This statement is misleading because the same page includes a section about configuring synchronous streaming replication. To avoid confusion, I think it makes sense to specify that record-based log shipping can be configured as either asynchronous or synchronous. Link: https://www.postgresql.org/docs/current/warm-standby.html
On Thu, 2025-08-21 at 15:20 +0000, PG Doc comments form wrote: > Page: https://www.postgresql.org/docs/17/warm-standby.html > > The documentation page about Log-Shipping Standby Servers after describing > that there are file-based log shipping and record-based log shipping > (streaming replication) states: "It should be noted that log shipping is > asynchronous, i.e., the WAL records are shipped after transaction commit.". > This statement is misleading because the same page includes a section about > configuring synchronous streaming replication. To avoid confusion, I think > it makes sense to specify that record-based log shipping can be configured > as either asynchronous or synchronous. I think that the statement you quote is not only misleading, but wrong. WAL can get shipped before the transaction commits. Perhaps the sentence had better be It should be noted that by default, log shipping is asynchronous, i.e., the primary server does not wait until the standby receives the data. Yours, Laurenz Albe
On Mon, 2025-08-25 at 09:58 +0200, Laurenz Albe wrote: > On Thu, 2025-08-21 at 15:20 +0000, PG Doc comments form wrote: > > Page: https://www.postgresql.org/docs/17/warm-standby.html > > > > The documentation page about Log-Shipping Standby Servers after describing > > that there are file-based log shipping and record-based log shipping > > (streaming replication) states: "It should be noted that log shipping is > > asynchronous, i.e., the WAL records are shipped after transaction commit.". > > This statement is misleading because the same page includes a section about > > configuring synchronous streaming replication. To avoid confusion, I think > > it makes sense to specify that record-based log shipping can be configured > > as either asynchronous or synchronous. > > I think that the statement you quote is not only misleading, but wrong. > WAL can get shipped before the transaction commits. Perhaps the sentence > had better be > > It should be noted that by default, log shipping is asynchronous, i.e., > the primary server does not wait until the standby receives the data. Here is a patch for that. Yours, Laurenz Albe
Вложения
On Wed, Aug 27, 2025 at 02:13:21PM +0200, Laurenz Albe wrote: > Here is a patch for that. > --- a/doc/src/sgml/high-availability.sgml > +++ b/doc/src/sgml/high-availability.sgml > @@ -527,8 +527,8 @@ protocol to make nodes agree on a serializable transactional order. > </para> > > <para> > - It should be noted that log shipping is asynchronous, i.e., the WAL > - records are shipped after transaction commit. As a result, there is a > + It should be noted that log shipping is asynchronous, i.e., the primary server does > + not wait until the standby receives the data. As a result, there is a > window for data loss should the primary server suffer a catastrophic > failure; transactions not yet shipped will be lost. The size of the > data loss window in file-based log shipping can be limited by use of the Yep, the original statement is rather inexact. Now, your new wording does not make me really comfortable with the case of cascading stanbys in scope, because the asynchronous property applies to them all the time. Hmm. I'd suggest to use a simpler reformulatione, like this one to outline that there is no relationship between the timing of a transaction commit and the timing where the commit records are flushed on a standby server: It should be noted that log shipping is asynchronous, i.e., the WAL records may be shipped after transaction commit. -- Michael
Вложения
On Mon, Sep 1, 2025 at 1:20 AM Michael Paquier <michael@paquier.xyz> wrote: > > Yep, the original statement is rather inexact. Now, your new wording > does not make me really comfortable with the case of cascading stanbys > in scope, because the asynchronous property applies to them all the > time. This is another unclear part. As I understand in configuration `Master -> Upstream -> Downstream` replication between Master And Upstream still can be synchronous, while between Upstream and Downstream is't always async. Am I wrong here? -- Artem Gavrilov Senior Software Engineer, Percona artem.gavrilov@percona.com
On Mon, 2025-09-01 at 08:20 +0900, Michael Paquier wrote: > On Wed, Aug 27, 2025 at 02:13:21PM +0200, Laurenz Albe wrote: > > Here is a patch for that. > > --- a/doc/src/sgml/high-availability.sgml > > +++ b/doc/src/sgml/high-availability.sgml > > @@ -527,8 +527,8 @@ protocol to make nodes agree on a serializable transactional order. > > </para> > > > > <para> > > - It should be noted that log shipping is asynchronous, i.e., the WAL > > - records are shipped after transaction commit. As a result, there is a > > + It should be noted that log shipping is asynchronous, i.e., the primary server does > > + not wait until the standby receives the data. As a result, there is a > > window for data loss should the primary server suffer a catastrophic > > failure; transactions not yet shipped will be lost. The size of the > > data loss window in file-based log shipping can be limited by use of the > > Yep, the original statement is rather inexact. Now, your new wording > does not make me really comfortable with the case of cascading stanbys > in scope, because the asynchronous property applies to them all the > time. > > Hmm. I'd suggest to use a simpler reformulatione, like this one to > outline that there is no relationship between the timing of a > transaction commit and the timing where the commit records are flushed > on a standby server: > It should be noted that log shipping is asynchronous, i.e., the WAL > records may be shipped after transaction commit. That is a less invasive change and probably preferable. The attached patch does it like you suggested. I noticed that the paragraph speaks about the asynchronicity of replication and the potential of data loss, so I couldn't resist the temptation to add a remark that synchronous streaming replication can avoid that problem. Yours, Laurenz Albe
Вложения
On Mon, 2025-09-01 at 13:51 +0200, Artem Gavrilov wrote: > As I understand in configuration `Master > -> Upstream -> Downstream` replication between Master And Upstream > still can be synchronous, while between Upstream and Downstream is't > always async. Am I wrong here?
On Mon, 2025-09-01 at 13:51 +0200, Artem Gavrilov wrote: > As I understand in configuration `Master > -> Upstream -> Downstream` replication between Master And Upstream > still can be synchronous, while between Upstream and Downstream is't > always async. Am I wrong here? I don't quite understand. Sure, you can have synchronous replication between the master and upstream. It is the "isn't always async" part that confuses me. Do you mean that WAL can reach downstream before the master commits? That is certainly the case. Yours, Laurenz Albe
Oh, sorry I made a typo, it should be "is always async". I was referring to this statement in docs about cascading replication: "Cascading replication is currently asynchronous". It sounds to me like the whole replication setup is async (M -> U ->D), but it's only the (U -> D) part that is always async. But probably it's a topic for another thread. My original problem was with the first sentence "It should be noted that log shipping is asynchronous". I think your original suggestion "It should be noted that by default, log shipping is asynchronous" sounds good as it highlights from the beginning that there is some variety. On Tue, Sep 2, 2025 at 9:34 AM Laurenz Albe <laurenz.albe@cybertec.at> wrote: > > On Mon, 2025-09-01 at 13:51 +0200, Artem Gavrilov wrote: > > As I understand in configuration `Master > > -> Upstream -> Downstream` replication between Master And Upstream > > still can be synchronous, while between Upstream and Downstream is't > > always async. Am I wrong here? > > I don't quite understand. Sure, you can have synchronous replication > between the master and upstream. It is the "isn't always async" part > that confuses me. Do you mean that WAL can reach downstream before > the master commits? That is certainly the case. > > Yours, > Laurenz Albe -- Artem Gavrilov Senior Software Engineer, Percona artem.gavrilov@percona.com
On Tue, 2025-09-02 at 11:22 +0200, Artem Gavrilov wrote: > My original problem was with the first sentence "It should be noted > that log shipping is asynchronous". I think your original suggestion > "It should be noted that by default, log shipping is asynchronous" > sounds good as it highlights from the beginning that there is some > variety. Hm, yes, we could add "by default". Yours, Laurenz Albe
On Tue, Sep 2, 2025 at 8:48 AM Laurenz Albe <laurenz.albe@cybertec.at> wrote: > > On Tue, 2025-09-02 at 11:22 +0200, Artem Gavrilov wrote: > > My original problem was with the first sentence "It should be noted > > that log shipping is asynchronous". I think your original suggestion > > "It should be noted that by default, log shipping is asynchronous" > > sounds good as it highlights from the beginning that there is some > > variety. > > Hm, yes, we could add "by default". > I think the issue here is that this section is supposed to focus on continuous archiving / file based WAL shipping, which is asynchronous. All of the complexity that is being discussed in this thread is really about WAL streaming, which IMO should not be discussed here. Per the docs, "Record-based log shipping is more granular and streams WAL changes incrementally over a network connection (see Section 26.2.5)." I actually think the thing that is wrong (or at least confusing) in the docs is this line "Directly moving WAL records from one database server to another is typically described as log shipping." because it is too loose with its definition. I don't recall postgres people referring to streaming replication as "wal shipping", that term is pretty exclusively used for continuous archiving. If you look in the aforementioned 26.2.5. Streaming Replication, the term "shipping" is only ever used in conjunction with the phrase "file-based log shipping". So with that said, I would suggest fixing this by changing the first sentence of paragraph 4 to "It should be noted that file based log shipping is asynchronous", as this also emphasizes that this section is focused on file based wal shipping. A larger fix would likely involve reworking this section to start with defining log shipping and how it is used in Postgres, and then continuing with the file based specific info (something like moving the third paragraph to the beginning and then editing things for clarity / readability). I could work up a patch for that if people were interested. Robert Treat https://xzilla.net
On Tue, Sep 02, 2025 at 11:10:42AM -0400, Robert Treat wrote: > So with that said, I would suggest fixing this by changing the first > sentence of paragraph 4 to "It should be noted that file based log > shipping is asynchronous", as this also emphasizes that this section > is focused on file based wal shipping. Not sure that there is a strong need for "file-based", still it is true that we could just remove the inexact part of the sentence and call it a day, as of: --- a/doc/src/sgml/high-availability.sgml +++ b/doc/src/sgml/high-availability.sgml @@ -527,8 +527,7 @@ protocol to make nodes agree on a serializable transactional order. </para> <para> - It should be noted that log shipping is asynchronous, i.e., the WAL - records are shipped after transaction commit. As a result, there is a + It should be noted that log shipping is asynchronous. As a result, there is a -- Michael
Вложения
On Tue, 2025-09-02 at 11:10 -0400, Robert Treat wrote:
> I think the issue here is that this section is supposed to focus on
> continuous archiving / file based WAL shipping, which is asynchronous.
> All of the complexity that is being discussed in this thread is really
> about WAL streaming, which IMO should not be discussed here. Per the
> docs, "Record-based log shipping is more granular and streams WAL
> changes incrementally over a network connection (see Section 26.2.5)."
Chapter 26.2. is "Log-Shipping Standby Servers".
The first line seems to confirm what you are saying:
Continuous archiving can be used to create a high availability (HA)
cluster configuration with one or more standby servers ready to
take over operations if the primary server fails. This capability
is widely referred to as warm standby or log shipping.
But one of the subsections is 26.2.5. "Streaming Replication", which
suggests that streaming replication is a kind of log shipping.
> I actually think the thing that is wrong (or at least confusing) in
> the docs is this line "Directly moving WAL records from one database
> server to another is typically described as log shipping." because it
> is too loose with its definition. I don't recall postgres people
> referring to streaming replication as "wal shipping", that term is
> pretty exclusively used for continuous archiving. If you look in the
> aforementioned 26.2.5. Streaming Replication, the term "shipping" is
> only ever used in conjunction with the phrase "file-based log
> shipping".
>
> So with that said, I would suggest fixing this by changing the first
> sentence of paragraph 4 to "It should be noted that file based log
> shipping is asynchronous", as this also emphasizes that this section
> is focused on file based wal shipping.
>
> A larger fix would likely involve reworking this section to start with
> defining log shipping and how it is used in Postgres, and then
> continuing with the file based specific info (something like moving
> the third paragraph to the beginning and then editing things for
> clarity / readability). I could work up a patch for that if people
> were interested.
I agree that it is a worthwhile goal to clarify the terms, and I
think that the whole chapter should be reorganized:
Sections 26.2.5. to 26.2.9. should be moved to a new chapter
26.3. "Streaming Replication" (which will renumber the present 26.3.
and 26.4.).
Perhaps "WAL shipping" would be a better term, with "WAL streaming"
as alternative.
But that would be a bigger endeavour that would require going over
bigger parts of the documentation. If you want to do that, I'd be
happy to review it.
But I think that the factually wrong statement that my patch
tries to address should get fixed first - who knows how long the
bigger patch would take.
I am OK with Michael's suggestion to just remove the wrong line,
although it wouldn't be bad to have an explanation of what we mean
by "asynchronous" here.
Yours,
Laurenz Albe
On Wed, Sep 03, 2025 at 09:37:08AM +0200, Laurenz Albe wrote: > I agree that it is a worthwhile goal to clarify the terms, and I > think that the whole chapter should be reorganized: > > Sections 26.2.5. to 26.2.9. should be moved to a new chapter > 26.3. "Streaming Replication" (which will renumber the present 26.3. > and 26.4.). I would not disagree with that, the situation in the docs can be confusing for one, as we mix file-based WAL files moved around and streaming with the replication protocol. One interesting portion is about replication slots, where we rely on XLogGetReplicationSlotMinimumLSN() to decide the retention threshold, Physical slots are updated in WAL senders via PhysicalConfirmReceivedLocation, meaning that the replication protocol is required. Mixing that with the file-shipping part is a mistake. Just moving the contents to a new "Streaming" section sounds like an improvement, but the "log-shipping" part would still suck. So this stands for cleanup as well, providing a better split. Perhaps we should embrace the term "file-based WAL shipping" or "file-based log shipping" and use that, giving a structure of: * WAL shipping methods, log-shipping methods or just "Log Shipping" ** File-based WAL shipping ** Streaming Warm standbys can use both methods. The part about planning, operation and preparing may be worth splitting outside the "method" portion.. The "continuous" archiving on standbys is not about streaming, but about the file-based method, so it would need to be inside the file-based subsection. We could replace "Log" with just "WAL", as well, if we're looking at more standardization of the whole area, while on it. > Perhaps "WAL shipping" would be a better term, with "WAL streaming" > as alternative. Perhaps that stands for improvement and more standarization. This term originates from 5e550acbc4d1 in 2006. The industry has changed a lot since and there may be standard terms which are much more adapted for the "modern" user, even if there's a lot of Postgres-ism in the architecture and how things are done. There have been some proposals, but nobody really stood up to commit something. > But that would be a bigger endeavour that would require going over > bigger parts of the documentation. If you want to do that, I'd be > happy to review it. > > But I think that the factually wrong statement that my patch > tries to address should get fixed first - who knows how long the > bigger patch would take. > > I am OK with Michael's suggestion to just remove the wrong line, > although it wouldn't be bad to have an explanation of what we mean > by "asynchronous" here. Yeah, this statement is confusing as-is because there is no dependency with the timing of a transaction commit, records may be shipped before or after depending on how your system balances your IO and/or CPU. I am not sure if this is worth applying on its own, TBH, because this stuff needs much more rework than a simple sentence. If somebody takes the time to write a patch, I'd be OK to step in this time for review and doing some reorganization of the whole section, even if that would mean a HEAD-only change. I had the attached staged at some point, for reference. Adding David Steele in CC, I recall that he may have done a proposal around all that for the docs, and he's involved in backrest. -- Michael